Dis cus si on Paper No. 16-002

How to Fill the Digital Gap? The (Limited) Role of Regulation

Wolfgang Briglauer, Carlo Cambini and Sauro Melani

Dis cus si on Paper No. 16-002

How to Fill the Digital Gap? The (Limited) Role of Regulation

Wolfgang Briglauer, Carlo Cambini and Sauro Melani

Download this ZEW Discussion Paper from our ftp server:

http://ftp.zew.de/pub/zew-docs/dp/dp16002.pdf

Die Dis cus si on Pape rs die nen einer mög lichst schnel len Ver brei tung von neue ren For schungs arbei ten des ZEW. Die Bei trä ge lie gen in allei ni ger Ver ant wor tung

der Auto ren und stel len nicht not wen di ger wei se die Mei nung des ZEW dar.

Dis cus si on Papers are inten ded to make results of ZEW research prompt ly avai la ble to other eco no mists in order to encou ra ge dis cus si on and sug gesti ons for revi si ons. The aut hors are sole ly

respon si ble for the con tents which do not neces sa ri ly repre sent the opi ni on of the ZEW.

How to Fill the Digital Gap?

The (Limited) Role of Regulation

Wolfgang Briglauer*, Carlo Cambini**, Sauro Melani***

December 2015

* Wolfgang Briglauer, Centre for European Economic Research (ZEW Mannheim),

P.O. Box 103443, 68034 Mannheim, Germany. MaCCI Mannheim Centre for

Competition and Innovation. E-mail: briglauer@zew.de, phone: +49 (0) 621 1235-

279.

** Carlo Cambini, Politecnico of Torino, Corso Duca degli Abruzzi, 24, 10129 Torino,

Italy. IEFE-Bocconi. Florence School of Regulation – European University Institute.

E-mail: carlo.cambini@polito.it, phone: + 39 (0) 110907292-7292.

*** Sauro Melani, Politecnico of Torino, Corso Duca degli Abruzzi, 24, 10129 Torino,

Italy. E-mail: sauro.melani@studenti.polito.it.

JEL: H5, L38, L43, L52

Key words: Next generation broadband networks; regulation; investment; adoption;

take-up; Digital Agenda Europe

Abstract

This paper provides evidence on the migration from an “old” technology to a “new”

technology, taking into account the impact that regulatory interventions on the old

one might have on the incentives to invest and adopt the new one. This analysis has

been applied to a sample of EU27 countries using panel data from 2004 to 2014 on

the adoption, coverage and take-up rate of ultra-fast broadband infrastructures,

whose development is one of the flagship initiatives of the Europe 2020 programmes.

Results show that a 1% increase in the regulated price to access the old technology

increases the adoption and the investment on the new broadband technology by

~0.45% and ~0.47%. These effects are not homogeneous across countries and are

weakened in Eastern European countries, where the existing old broadband

infrastructures are less developed than in the rest of Europe. It has also been shown

that the access price to old networks negatively affects the take-up rate of the new

technology-based services, thus calling for the need of more specific and

complementary demand side policy incentives to enhance service adoption.

~ 1 ~

1 Introduction

In a time of increasing digitalization, such as the one we are currently observing,

operators of “old” (copper-wire and coaxial cable based) broadband networks are

facing a huge increase in demand for bandwidth and real time criteria, due to the

presence of interactive multimedia services such as streamed video on demand, file

sharing, online gaming, and high definition television, as well as specific business

applications, such as cloud computing services or video conferencing. As a

consequence, the fibre-based deployment of ultra-fast broadband networks (”new” or

“Next Generation Networks” – NGNs) that enable a massive increase in bandwidth

capacity has become a major issue for regulators and telecom companies. The latter,

however, have to sustain costly investments to upgrade the infrastructure, which is

also fraught with high uncertainties as regards the future demand and regulatory

policies. At the same time, NGNs can be considered as a general purpose

technology (Bresnahan & Trajtenberg, 1995), which has the potential to trigger

productivity gains and growth across major economic sectors, such as health,

electricity and transport, on a massive scale.1

In view of the expected externalities that are involved, the European Commission

(EC) has decided to strengthen the competitiveness of Europe’s economy by

explicitly focusing on digital infrastructure and communication technologies. In order

to reach the related growth and productivity potential of NGNs, the Digital Agenda for

Europe (DAE) has specified goals in terms of network coverage and service

adoption: The DAE “seeks to ensure that, by 2020, (i) all Europeans will have access

to much higher internet speeds of above 30 Mbps and (ii) 50% or more of European

households will subscribe to internet connections above 100 Mbps” (European

Commission, 2010:19).2 While target (i) refers to a coverage level of 100 per cent of

the population, target (ii) refers to a minimum household adoption level. However,

recent market data (European Commission, 2014) have shown that both targets are

unlikely to be met, unless substantial infrastructure investments are introduced in the

coming years. Similar targets can be found in other jurisdictions outside the EU, such

as the “National Broadband Network” and the “Ultra-Fast Broadband Initiative“ in

Australia and New Zealand, respectively, the “Digital Divide Closing Plan” in South-

Korea or the “Connecting America: The National Broadband Plan” in the US.

What are the main drivers of ultra-fast broadband adoption and coverage in Europe?

What is the role that the existing regulation can play on the old “legacy” (“copper-

wire”) network to foster NGN adoption and coverage? In this paper, an attempt has

1 Numerous studies support the view that investment in (old) broadband infrastructures creates

positive effects on the economic system and leads to an increase in GDP growth (e.g. Röller & Waverman, 2001; Czernich, Falck, Kretschmer & Wößmann, 2011). In particular, Czernich et al. (2011) have shown that a 10% increase in the broadband adoption rate in OECD countries results in a 1-1.5% increase in the annual GDP per-capita. 2

The DAE is one of the seven flagship initiatives under Europe 2020. For further details about the DAE, the reader can refer to the European Commission’s website: http://ec.europa.eu/digital-agenda/digital-agenda-europe.

~ 2 ~

been made to answer these questions. Using a recent EU27 panel data set for the

years 2004 to 2014, and static and dynamic model specifications, the present work is

the first that has simultaneously examines the determinants of NGN coverage, NGN

adoption and the NGN take-up rate. The latter measure relates NGN adoption to

NGN coverage. The role of the EU regulatory policies, as embedded in the sector-

specific framework of electronic communication markets, is examined, as well as the

related market conditions, including relevant forms of competition within fixed

broadband markets (“intramodal”) and from mobile networks (“intermodal”),

deployment costs and demand characteristics. The market conditions in most of the

European countries so far appear to be insufficient to trigger the broad-scale

deployment of NGN. Accordingly, the focus of this work has been on examining

regulatory policies more closely, in particular, the so-called “unbundling” price, which

is the most relevant policy instrument in terms of incentivizing migration to NGNs

pertaining to investment and adoption. Unbundling prices are set directly by national

regulatory authorities (NRAs) in individual member states subject to framework

directives at the EU level (European Commission, 2000; European Commission,

2002a; European Commission, 2002b). In view of the dual DAE policy goals, and in

order to avoid inefficient NGN deployment, it is essential to identify the right

regulatory policies. Therefore, the present paper examines how relevant broadband

market regulations have an impact on both input-related NGN investment and output-

related NGN adoption, as well as their simultaneous impact on NGN take-up.

The European policy goals are closely interrelated, since investment in NGN, i.e.

network coverage, also depends on the (expected) adoption, i.e. (future) demand,

which in turn is determined by the attractiveness of specific NGN services and

applications. Only if consumers consider NGN services attractive enough, in terms of

innovations or quality improvements compared with old broadband services, will they

migrate to NGN. In this perspective, the take-up rate, is a useful indicator of the

willingness of consumers to migrate to the new infrastructure. The more consumers

are satisfied with conventional broadband services, or the more consumers are

reluctant to adopt new technologies, the greater the gap will be with the newly

installed network capacity. A high take-up rate, with adoption being close to capacity

in terms of NGN coverage, avoids social costs due to over-capacities. For these

reasons, the analysis also focuses on the NGN take-up rate, because of its primary

role in the EU scenario, by empirically assessing its main determinants. In this

perspective, this paper is the first to attempt to empirically assess the complex

interplay between regulation on an old technology and investment and adoption of a

new technology, as recently proposed in a theoretical framework by Bourreau,

Cambini and Dogan (2012) and Bourreau, Cambini and Dogan (2014).

Results show that the access price imposed on the old legacy infrastructure

significantly affects both NGN adoption and coverage. In particular, results show that

a 1% increase in the unbundling price increases NGN adoption and NGN investment

by ~0.45% and ~0.47%, respectively. This implies that a policy measure that

increases the cost of accessing the old broadband infrastructure, though affecting

~ 3 ~

competition, could exert a positive effect on incentivizing the deployment of a new

fibre infrastructure (hence, expanding NGN coverage), but also on the adoption of

the new connections, by reducing the gap between the retail prices between old and

new technology based broadband services. However, and interestingly from a policy

perspective, these effects are greatly reduced in Eastern European countries that are

characterized by a lack of a well-developed legacy infrastructure: when controlling for

this heterogeneity across countries, it has been found that the role of the unbundling

regime is offset in Eastern European countries. This result casts doubts on the EC´s

current policy of creating a single market in Europe with uniform regulatory rules to

be applied in all countries. Clearly, the possible changes in the unbundling prices are

only relevant in certain EU countries (mostly EU15), but not over the entire continent.

Finally, the take-up rate estimations results have shown that increasing the price of

the access price decreases the take-up rate, since adoption increases but less than

proportionally to coverage. From a policy perspective, this implies that using a single

instrument (i.e. the price for local loop unbundling, LLU) to influence both demand

adoption and coverage is not enough, and other instruments are needed to support

demand adoption, such as vouchers or tax deductions.

The remainder of the paper is organised as follows: Section 2 reviews the NGN

related literature, focusing in particular on empirical literature. Section 3 describes the

basic hypotheses concerning the relationship between regulation and competition on

NGN coverage, adoption and take-up. Section 4 outlines the panel dataset that

underlies the empirical examination. Section 5 presents the empirical baseline

specifications and the related econometric issues. Section 6 describes and interprets

the main results. Section 7 summarises and compiles the most relevant trade-offs for

policy makers.

2 Literature review

The economic literature on the migration from old to new broadband technology is

relatively recent, and evidence on this phenomenon is relatively scant.

The deployment of fibre infrastructures does not immediately replace copper or cable

legacy networks, suggesting that the transition from old infrastructures to new

infrastructures will go slowly. This implies that, during a transition phase, two different

infrastructures will operate in parallel, and presumably each type of network will be

regulated with a different set of rules. The incentives to invest in fibre infrastructures

will therefore also be influenced by the terms of access set for the legacy copper

networks.3 The recent theoretical literature (Bourreau et al., 2012; Bourreau et al.,

2014; Inderst & Peitz, 2012) has focused on how access regulations on an existing

old network affect infrastructure investments in new networks and favour the

migration, at a retail level, from the old to the new broadband infrastructure.

3 It should be noted that cable coax networks also constitute old broadband networks. However, only

copper-wire based (“legacy”) networks have been subjected to sector-specific regulations, such as unbundling, in the EU regulatory framework for electronic communications markets.

~ 4 ~

The related empirical literature on NGN investment (coverage) is relatively scant.

Minamihashi (2012) has examined whether unbundling regulations imposed on the

Japanese incumbent operator have prevented entrants from self-deploying new

broadband infrastructures, using municipal level data from 2005 to 2009. The author

has found that unbundling regulations hinder entrants from investing in their own

NGN infrastructure. However, during the analysed years, the incumbent’s NGN

investments were not hindered by the unbundling regulations. Bacache, Bourreau

and Gaudin (2014) have examined the incentives embedded in the EU regulatory

framework on migration from old to new broadband infrastructures using biannual

data from 15 European member states over a period from July 2002 to July 2010.

The authors related the number of broadband lines based on new infrastructure to

the number of unbundling lines and found that unbundling regulations did not foster

entrants investing in NGN. Briglauer (2015) has examined the impact of broadband

regulations, including the unbundling price, on NGN investment, utilizing EU27 panel

data from 2004 to 2013. The author has found that, as the unbundling price

increased, so did the average incentives for NGN investment.

As far as NGN adoption is concerned, the existing empirical literature presents (i)

several contributions related to old broadband markets, but only (ii) a few NGN-

related publications. Regarding point (i), several relatively old papers exist that have

dealt with the determinants of broadband adoption in both the US and European

countries. Bouckaert, van Dijk and Verboven (2010) have examined the determinants

of broadband adoption from 2003 to 2008 in OECD countries and have found that

infrastructure-based competition has a positive impact on broadband adoption. The

first paper to use EU data was that of Distaso, Lupi and Maneti (2006), who found

that infrastructure-based competition was the main driver of broadband adoption and

that it played a more important role than service-based competition, especially in the

longer term. More recently, Nardotto, Valletti and Verboven (2015) have employed

disaggregated broadband data related to the old telecom infrastructure in the UK for

the period December 2005 to December 2009. The authors have shown that

unbundling in the UK has not resulted in an increase in broadband adoption but has

positively affected service quality.

The above mentioned papers have shed some light on the impact of infrastructure-

based competition and access regulation on standard broadband adoption. However,

although they are interesting, they are of limited interest for a better understanding of

NGN adoption, where the presence of a relatively good legacy infrastructure may

represent a constraint to the development of NGN adoption. There are very few

papers that deal with NGN demand adoption (point (ii)). Wallsten and Hausladen

(2009) have estimated the effects of broadband regulations on NGN adoption with

data from EU27 countries from 2002 to 2007, thus covering the very early market

phase. They have found that countries where unbundling is more effective

experience lower NGN adoption. In their paper, the authors only examined the

presence of unbundling regulation, but did not provide any evidence on the possible

impact of the price of unbundling access on NGN adoption. Samanta, Martin, Guild

~ 5 ~

and Pan (2012) have examined the demand-side determinants of high-speed

broadband deployment using International Telecommunication Union (ITU) and

OECD data for 25 countries (for the years) from 1999 to 2009. The authors employed

a dummy variable to capture the extent of unbundling regulation and found that this

variable had no significant impact. More recently, Briglauer (2014) has investigated

the determinants of NGN adoption for EU27 member states from 2004 to 2013. The

author has found that the more effective the previous broadband access regulation

was, the more negative the impact on adoption. He also found that competitive

pressure from mobile networks affects adoption in a non-linear manner.

It should be pointed out that none of the above papers analysed the cross price effect

of an old network, i.e. a change in the local loop unbundling price on NGN adoption.

This type of analysis can be considered extremely important since, as the theoretical

models show, the consumers’ migration, at the retail level, from old broadband

connections to fibre-based connections depends on the relative price difference

between the NGN retail services and the standard broadband ones. In fact, when the

access price of the legacy network is low, the retail prices for the services that rely on

this network are also low. Hence, in order to encourage customers to switch from the

legacy network, operators would need to introduce low-priced NGN services.

Furthermore, none of the existing empirical studies has analyzed the determinants of

the NGN take-up rate. As mentioned in the introduction, the latter is an important

indicator of consumer willingness to adopt new services, of capacity utilization and of

the extent to which policy targets are achieved.

Overall, the present paper has the aim of examining the potential role of regulation

on stimulating both policy goals, i.e., coverage and adoption. In order to provide

useful information to the policy debate, the impact on NGN coverage and NGN

adoption is estimated as well as the impact on the NGN take-up rate in separate

regressions.

3 Hypotheses

As outlined in the introduction, the key policy variable of interest is the regulated

wholesale access price to the old (legacy) infrastructure, i.e. the local loop

unbundling price. The current policy debate is focused on how to revise the

regulation of this wholesale price in order to foster both ultra-fast broadband

coverage and adoption by end users. In fact, the EC is currently modifying the

regulatory framework in order to fulfill the EU targets defined within the DAE

program.4 The present analysis thus focuses on this key variable.

4 The reader can refer to the relevant recommendations of the European Commission related to

regulated access to Next Generation Access Networks (2010/572) and non-discrimination and costing methodologies (2013/466), as well as to the current public consultation on the review of the regulatory framework for electronic communication networks and services (information available at: http://ec.europa.eu/digital-agenda/en/news/public-consultation-evaluation-and-review-regulatory-framework-electronic-communications).

~ 6 ~

To this aim, it is important to derive sound testable hypotheses from the recent

theoretical literature on the economics of technology migration. The first systematic

theoretical analysis on this issue was provided by Bourreau et al. (2012) and

Bourreau et al. (2014). The authors consider a model in which access to the legacy

copper network is available throughout an entire country, and an incumbent that is

subject to access regulation on the old network, and an (unregulated) entrant

operator competes for the provision of retail broadband services to consumers by

investing in a new ultra-fast broadband infrastructure. The entrant operator could also

demand access in the form of LLU. Their main results show that NGN coverage

varies non-monotonically with the LLU access price. This result is due to the

coexistence of three different effects: (i) the “replacement effect”, which hinders

infrastructure investment by alternative operators when the access price is low; (ii)

the “wholesale revenue effect”, which discourages the incumbent from investing in a

higher quality network when the access price is high (since the entrant may invest in

reaction, and the incumbent will then lose some of its wholesale profits); and finally

(iii) the “business migration effect”: when the LLU access price is low, the retail prices

of the services that rely on the copper network are also low. Therefore, in order to

encourage customers to switch from an old to NGN services, operators should also

offer low prices for the NGN services. This effect reduces the profitability of the NGN

infrastructure, and hence, the incentives to invest in it.

From this analysis, four different testable hypotheses can be drawn. First, the effect

of the access price on the investment in the new technology networks is in general

ambiguous. As pointed out, three effects are at play and the aggregate NGN

coverage generally varies non-monotonically with the access price of the copper

network. This nonlinear effect emerges mainly from the so-called wholesale revenue

effect (Bourreau et al., 2012), which discourages the incumbent from investing in a

higher quality network when the access price is high in order not to jeopardize the

extra-return they can obtain from providing access to their old infrastructures to third

parties. In other words, while increasing access to old networks would incentivize not

only the entrants but also the incumbents to invest in new infrastructures, and would

favour the consumers to switch to adopting the new technology based services, the

extra-return on the old legacy infrastructure would limit the incentives of the

incumbents to invest. This implies that the effect of an increase in the access price is

not clear a priori, unless it would be possible to control for the wholesale revenue

effect; if this were possible, the following testable hypothesis would emerge:

H1: Assuming that it is possible to control the wholesale revenue effect,

an increase in the regulated access price to the old technology would

boost the new technology investment and expand its coverage.

Second, regarding NGN adoption, an issue that is extremely important and that may

affect consumer migration, at a retail level, from the standard copper infrastructures

to NGN connections, is the relative price difference between the NGN retail services

and the standard broadband ones. Indeed, when the access price on the legacy

~ 7 ~

network is low, the prices of the services that rely on this network are also low.

Hence, in order to encourage customers to move away from the legacy network,

operators would need to introduce low-priced NGN offers. The latter effect, which is

referred to as the business migration effect, implies that the access price on copper

networks (i.e. the LLU price) may have a considerable effect on NGN adoption:

assuming that the retail market for copper-based broadband services is substantially

competitive, any increase in the cost of LLU prices would be translated into a higher

cost of the basic broadband connections, thus making it less attractive than the NGN-

based services. The following can therefore be tested:

H2: An increase in the regulated access prices to the old technology

would make old broadband services similar to the new technology based

services, and as a result the adoption of the latter would increase.

An analysis of the take-up rate, which relates NGN adoption to NGN coverage seems

less insightful because the adoption and coverage of the new infrastructures, as

tested in the previous hypotheses, are simply being compared. However, as pointed

out in the main Introduction, this index is not only a useful indicator of the willingness

of consumers to migrate to a new infrastructure, but also the key policy variable used

by the EC to define specific targets, in terms of NGN adoption and coverage, and

implicitly also to define the take-up rate target. As shown in Hypotheses 1 and 2, it

could be expected – after controlling for specific effects – that both adoption and

coverage would be positively affected by an increase in the access price to the old

networks and therefore the expected effect of this price on the take-up rate is ex ante

indeterminate. The following can therefore be tested:

H3: The impact of an increase in the regulated access price to the old

technology on the take-up rate of the new technology depends on the

incremental effect of such a price increase on the new technology

adoption rate and coverage.

Finally, the above hypotheses hold for countries in which the old legacy infrastructure

is well established on a nation-wide scale. In those countries, the access price to this

infrastructure plays a relevant role. However, in countries where the legacy network

is not very well developed, mostly for historical reasons, the role of the access price

on the new technology coverage and adoption should be weaker. The following

hypothesis emerges:

H4: In countries in which the presence of the old legacy technology is

limited, an increase in the access price to the old technology should play

a minor role in incentivizing the coverage and the adoption of the new

technology.

~ 8 ~

4 Data

In the empirical analysis, country level panel data for EU27 member states from 2004

to 2014 have been considered. The data have been gathered from several different

sources: FTTH Council Europe5 provides annual NGN coverage and adoption data

from 2004 to 2014, thus covering almost the entire period of NGN deployment in EU

member states. NGN coverage and adoption data also form the basis of the NGN

take-up rate and the NGN gap measure, as discussed in section 4.1. Owing to the

fact that some values are missing, there are fewer observations than the maximum

number of 297 (27×11).6 Furthermore, any unrealistically high take-up rates ≥ 0.75

which have occasionally been observed at the beginning of the NGN deployment in

Spain, Poland, Slovenia and the United Kingdom from 2004 up to 2006, as well as in

Greece from 2007 up to 2010 have been dropped. Generally, it is possible to

observe, from Figures A.1 and A.2, that NGN take-up rates are significantly higher at

the very beginning of NGN deployment, i.e. in the years from 2004 to 2007. One

obvious explanation might be that NGN were initially deployed in areas in which there

was a very high demand (e.g. universities, public administrations, large businesses,

residential consumers with high willingness to pay), which resulted in a high adoption

of installed NGN connections and hence in high take-up rates. Furthermore, during

the first years of NGN deployment, many field experiments were conducted by

operators in which the consumers were either volunteers or they obtained special

offers (in some cases without having to pay any extra price). Hence, a very high take-

up rate can be observed with respect to the selected and targeted consumers, which,

at the same time constituted a substantial segment of the market.

As regards the independent variables, the EU Digital Agenda Scoreboard7 provides

yearly data on broadband regulations. As for the competition variables, the data on

intermodal competition from mobiles (“wireless”) and intramodal broadband

competition (“wireline”) have been provided by Euromonitor,8 the International

Telecommunications Union (ITU)9 and the EU Digital Agenda Scoreboard.

Euromonitor also provides data on the number of households and on the Networked

5 These data are available to FTTH Council Europe members at:

http://www.ftthcouncil.eu/resources?category_id=6. 6 There is basically no data for Malta and Cyprus on NGN deployment for the entire period of interest

and these countries have therefore been excluded. Data on NGN coverage are also missing for the Czech Republic, Germany, Estonia, Poland, Slovenia and the United Kingdom in 2004, for Latvia, Lithuania, Portugal, Romania and Slovakia in 2004 and 2005, and for Greece, Luxembourg, Hungary and Bulgaria for a time span of up to five years starting from 2004. 7 The EU “Digital Agenda Scoreboard” is available at:

http://ec.europa.eu/information_society/policy/ecomm/library/communications_reports/index_en.ht. Values are missing for Bulgaria for the years from 2003 to 2006, for Romania from 2003 to 2004, for Estonia for 2003 and 2012, as well as for Cyprus, Latvia, Lithuania, Hungary, Malta, Poland, Slovenia and Slovakia for 2003. 8 The Euromonitor International database is commercially available at: http://www.euromonitor.com/.

Telecommunication revenue values are missing for the Netherlands for the year 2003, for Greece for 2013 and for Romania and Slovenia for both 2003 and 2004. 9 The ITU World Telecommunication/ICT Indicator Database is available at: http://www.itu.int/ITU-

D/ict/statistics/.

~ 9 ~

Readiness Index. Eurostat10 provides data on education and ICT labour costs.

Market Line11 provides data on the percentage of urban population and population

density. Finally, World Bank12 and the International Monetary Fund13 provide data on

GDP, and the European Central Bank has provided data on the long-term interest

rates.14 All the independent variables (discussed in section 4.2) are available for the

years from 2003 to 2013. As the data availability differs according to the variable, an

unbalanced panel data set has been employed.15

The variable descriptions are listed in the Annex in Table A.1 and Table A.2,

respectively, together with the data sources and summary statistics.

4.1 Dependent variables

NGN coverage, NGN_cov, measures the total number of deployed lines normalized

to the total number of households (“homes passed”). Network coverage thus

represents the installed capacity, in physical units, where the term “homes passed”

refers to the number of consumers with potential access to NGN infrastructure. On

the other hand, the variable NGN adoption, NGN_adop, measures the total number

of consumers (normalized to households) who subscribe to at least one service

offered via the NGN connection on a commercial basis (“homes connected”).

The NGN take-up rate, NGN_tur, is the ratio between NGN adoption and NGN

coverage, and thus ranges continuously in the [0;1] interval, as adoption cannot be

higher than the installed capacity. In the case of optimal network utilization, the

variable takes on the value of one. However, the denominator of NGN_tur, i.e.,

NGN_cov, is not in the [0;1] interval, as household coverage is already above 100%

in some member states. This is due to a parallel coverage with the NGN

infrastructure, in particular in urban areas, where homes are supplied with both cable

and traditional telecommunication operators. However, the saturation level for NGN

adoption is 100%, as households normally will not subscribe to multiple connections,

considering the huge bandwidth capacity of a single NGN connection. In order to

capture this asymmetry in maximum adoption and coverage levels, an alternative

take-up measure has been defined as a robustness variable, that is, NGN_gap,

which indicates the difference between NGN coverage and NGN adoption in absolute

terms, where the upper bound of the variable NGN_cov is set equal to one.

10

Data are available at: http://epp.eurostat.ec.europa.eu/portal/page/portal/information_society/-data/database. A few values are missing for the Austrian, Italian and Swedish dwelling permits,, as well as for the labour cost variables for Ireland and Greece. Networked Readiness Index values are also missing for Malta and Cyprus for 2003 and for Romania for 2006. Values pertaining to the number of internet users for Greece are missing (for the years) from 2011 to 2013. 11

Data are commercially available at: http://advantage.marketline.com/PageForbidden?returnUrl=%2F. 12

The World Bank’s “World Development Indicators” are available at: http://data.worldbank.org. 13

Data are available at: http://www.imf.org/external/data.htm. 14

Values are missing for the long-term interest rate of Romania (for the years) from 2003 to 2005. 15

In addition, there are some gaps in the raw data and the corresponding missing data had to be linearly interpolated. Overall, ~0.8% of all the raw data were calculated using linear interpolation or had to be extrapolated constantly for the future.

~ 10 ~

NGN coverage and adoption rates follow an investment adjustment and diffusion

process, respectively, as evidenced by the related empirical literature, Figure A.1,

pertaining to EU27 countries, and the two sub-groups of EU15 and Eastern

European Countries16 in Figure A.2, show that the NGN take-up rate does not follow

a specific growth pattern but instead fluctuates around average mean values

throughout most of the analysis period. It should also be noted that the mean of the

NGN take-up rate is well below the target take-up rate implied by the DAE (i.e., 0.5).

The graphical evidence suggests that, while coverage seems to present a rather

similar trend across EU countries, adoption and the take-up rate of NGN services are

larger in Eastern European countries where the presence of the old legacy

infrastructure is limited or even absent, and where any regulatory policies towards a

revision of the access prices to the legacy infrastructure appears to be less relevant.

4.2 Independent Variables

The independent variables can be divided into four categories: (i) regulation; (ii)

competition; (iii) controls; and (iv) time period- and country fixed effects.

(i) The monthly unbundling access price, measured in €, llu_price, is the most

relevant form of (wholesale) broadband regulation when considering migration from

old to new broadband networks, and which is set directly by NRAs. However, as

Bacache et al. (2014:205-206) pointed out, only a few unbundling price changes

were imposed by NRAs in the past, which makes identification of the overall effect

difficult. In order to circumvent this problem, an additional unbundling variable has

been introduced by referring to a measure that captures the effectiveness of the

unbundling regime (Briglauer, 2015). Accordingly, the variable, i_price_llu_sh,

combines the unbundling price, llu_price, with the respective unbundling market

share, ms_llu. The latter is bound between 0 and 1, where the upper limit indicates

that all the retail broadband connections are offered via unbundling. This variable

also provides a better representation of the overall complexity of unbundling regimes

which include several other institutional and technical regulations besides the

monthly access charge. Overall, both variables, llu_price and i_price_llu_sh have

been used as our main regulatory variables.

Furthermore, the variable sa_price is used as an instrumenting variable; this variable

represents the monthly cost of “shared access”, measured in €. Whereas unbundling

provides full LLU access to the incumbent´s access lines, shared access only

provides limited access to the upper line bandwidth. Accordingly, the regulated price

of shared access products represents approximately one half of the unbundling price

(see Table A.2). Hence, a change in shared access prices should not induce entrants

to switch to much more cost intense self-provision of the NGN infrastructure, whereas

the unbundling price – which represents the most investment intense business case

for entrants – has an impact on the entrant´s investment decision at the margin. At

16

Bulgaria, Czech Republic, Estonia, Latvia, Lithuania, Hungary, Poland, Romania, Slovenia and Slovakia have been included in the East European country group. Hence, the EU15 group includes all other EU27 member states, except Malta and Cyprus.

~ 11 ~

the same time, shared access and unbundling prices are closely related, since both

are determined by NRAs on the basis of (common) network costs. As the latter

represent about 60-80% of the total costs (ERG, 2007), our regulatory variables, in

particular the unbundling price, also represent a valid proxy of the average retail

broadband price.

(ii) Three variables are related to competition in retail broadband markets: the first

one stems from mobile networks (“intermodal” wireless competition). In order to

account for mobile network competition, the variable fms, which relates the total

number of mobile subscriptions to the total number of fixed landlines, has been used.

The second competition variable, bb_ne, represents the entrant's retail market share

in fixed broadband lines, and thus the impact of wireline (“intramodal”) competition on

old broadband markets on emerging NGN markets. Thirdly, the legacy variable

measures a country’s total stock of fixed-linked copper-wire connections, and it is

therefore able to directly capture the replacement effect for the incumbent that stems

from all wholesale and retail services of the incumbent´s old network infrastructure.

Accordingly, this variable also captures the wholesale revenue effect as defined in

Section 3.

(iii) A broad set of demand and cost controls, Z, have also been included, in line

with the previous empirical literature, and on the basis of industry knowledge (e.g.

FTTH Council Europe, 2013: 36-47). A detailed description of all the controls can be

found in Table A.1 in the Annex.

(iv) Finally, period effects, δ, and country fixed-effects, θ, have also been

considered. Including period effects makes it possible to control for relevant industry

developments that are common to all EU27 member states throughout the entire

period of analysis, such as different market phases or changes in equipment and

material prices, which are determined by industry standards and global markets. The

fixed effects are related to some of the main cost conditions, such as the topographic

and demographic characteristics. Likewise, supply- and demand-oriented NGN

subsidies, once having been determined by local or national governments, generally

stay in place for a longer period of time.

5 Empirical specifications

In view of the different diffusion patterns and the interdependencies underlying the

dependent variables, a two-fold research strategy has been employed: the empirical

baseline specifications for the separate regressions of the NGN investment and the

adoption models have been presented in section 5.1. Although any adoption and

investment process is inherently dynamic, the development of the NGN take-up rates

points to a static baseline specification, which has been outlined in Section 5.2. The

adopted estimation and identification strategy has been described in Section 5.3.

~ 12 ~

5.1 Dynamic NGN investment and adoption models

According to the ICT related empirical literature, the dynamics of the adoption

process, which is due to network effects or consumer inertia, as well as of the

investment process, which is mainly determined by the extent of adjustment costs,

can be captured by including the lagged dependent variable as an additional right-

hand side explanatory variable (Kiiski & Pohjola, 2002; Grajek & Röller, 2011). In

view of the discussion presented in sections 3 and 4, the dynamic reduced-form

models, in which NGN investment (superscript c denotes coverage in equation 1) and

NGN adoption (superscript a denotes adoption in equation 2) are expressed in logs17

for EU member state i and year t, read as follows:

(1)

itti

cc

t

c

i

c

ti

c

ti

cti

c

ti

cti

c

ti

c

ti

c

ti

cc

it

NGN

legacynebbnebbfmsfms

shllupriceipricelluNGN

)cov_ln(

__

)___ln()_ln()cov_ln(

)1(1)1(

)1(7)1(2

6)1(5)1(2

4)1(3

)1(2)1(10

´Z

(2)

itti

aa

t

a

i

a

ti

a

ti

ati

a

ti

ati

a

ti

a

ti

a

ti

aa

it

adopNGN

legacynebbnebbfmsfms

shllupriceipricelluadopNGN

)_ln(´

__

)___ln()_ln()_ln(

)1(1)1(

)1(7)1(2

6)1(5)1(2

4)1(3

)1(2)1(10

Z

The coefficients of the lagged dependent variables measure the constant speed of

diffusion )1( 1

a and the speed of adjustment )1( 1

c in the NGN adoption and

NGN investment specifications, respectively. The dynamic specifications are correct,

and give rise to an endogenous growth process if 0 < α1 < 1. Equation (1) and

equation (2) also depend on the main variables of interest, i.e., regulation, in terms of

the variables ln(llu_pricei(t-1)) and ln(i_price_llu_shi(t-1)), and competition, in terms of

the variables fmsi(t-1), bb_nei(t-1) and legacyi(t-1). In order to estimate the potential non-

linear relations as regards competition variables, squared terms of the variables,

related to intermodal (fmsi(t-1)) and intramodal (bb_nei(t-1)) competition, have also been

included in our baseline specifications (Schmutzler & Sacco, 2011). Furthermore, a

vector of controls, Zi(t-1), with demand controls and cost controls has been included in

the adoption and coverage baseline equations, respectively. Finally, εit and φit

represent additive error terms, θi´s country-specific effects and λt´s period effects.

17

A log transformation helps to stabilize the series of dependent variables and is also necessary to capture the dynamics of the data generating diffusion and adjustment processes adequately. In order to be able to interpret the main variables of interest in terms of elasticities, the variables related to the unbundling price have also been expressed as a logarithm in the dynamic specifications.

~ 13 ~

5.2 Static NGN Take-Up Rate model

As the take-up rate does not exhibit an endogenous adjustment process (Figure A.1),

modelling a static specification appears to be a reasonable choice. The empirical

baseline specification for the NGN take-up rate model, NGN_turit, for EU member

state i and year t, reads as follows:

(3)

it

tur

t

tur

i

c

ti

a

ti

tur

ti

turti

tur

ti

turti

tur

ti

tur

ti

tur

ti

turtur

legacynebbnebbfmsfms

shllupriceipricellu

)1()1(

)1(7)1(2

6)1(5)1(2

4)1(3

)1(2)1(1it

´

__

____NGN_tur0

ZZ

Equation (3) contains the same list of explanatory variables as in the dynamic

specifications, except for the lagged-dependent variable ( 01 tur ) and the fact that

equation (3) controls for both demand and cost shifters, a

ti )1( Z and c

ti )1( Z . It should be

noted that equations (1)-(3) include lagged values of all the explanatory variables in

order to employ the entire available data set (as described in section 4).18,19

5.3 Estimation and identification strategy

In order to identify causal effects, two-way fixed-effect regressions have been

employed to control for potential endogeneity due to unobserved and time-constant

heterogeneity at the country level (θ) as well as period effects (λ) to control for any

time specific shocks that are common to all cross-sectional units (member states).

However, estimating equations (1) and (2) by means of an ordinary fixed-effect

(least-squares-dummy-variable, LSDV) estimator, would yield inconsistent and

biased results, since the lagged dependent variable and the error terms that include

the fixed effects would be correlated (Nickell, 1981). In order to identify the

parameters of the dynamic models, a bias-corrected fixed-effect estimator (LSDVC),

developed by Bruno (2005a) and Bruno (2005b) specifically for dynamic unbalanced

panel data, and a small number of cross-sectional units (N = 25), has been

employed.

Second, by lagging all the explanatory variables, the dependent variables in

equations (1)-(3) are related to the pre-determined values of the independent

variables, which mitigates endogeneity due to time-variant heterogeneity if the model

is dynamically complete, i.e. in the absence of serial correlations. Although pre-

18

Moreover, it also makes sense to assume that adoption and investment decisions at a particular point in time do depend on the conditions of the latter period, in view of switching and adjustment costs on the side of consumers and operators, respectively. Investing firms are faced with rigidities related to the legal and institutional framework, as well as technical complexities of NGN deployment, and consumers of broadband services are usually subjected to long term contracts (up to two years) and non-transparent tariff structures. 19

For the sake of clarity, the indices have been dropped in the remainder of the paper.

~ 14 ~

determinedness, or sequential exogeneity, is in fact reasonable for dynamic

autoregressive models, such as those in equations (1) and (2) (Wooldridge, 2002:

299-300), serial correlation in the static specification (equation (3)) has to be

addressed in a different way. The nature of a serial correlation is first examined and

then the serial components are removed using a suitable data transformation.

Third, to rule out potential endogeneity due to reverse causality, Granger causality

tests (Granger, 1969) have been also performed. The results, which are reported in

Table A.3 in the Annex, indicate that there is no evidence of reverse causality.20

Fourth, a large number of demand and cost controls have been employed in order to

further reduce any remaining omitted variable bias that might be due to time-variant

heterogeneity.

Finally, as parts of the robustness specifications, the main regulatory variable,

llu_price, has been instrumented with the sa_price variable as well as with some

other exogenous cost shifters.

6 Empirical Results

According to the aforementioned two-fold research strategy, the results of the

dynamic models are first discussed in Section 6.1 and those of the static take-up rate

model are given in Section 6.2. The estimation results of the individual models on

NGN coverage and adoption also provide important information for the interpretation

of the estimation results pertaining to the take-up rate model. Finally, additional

estimations are presented in Section 6.3 in order to examine the robustness of the

main estimation results.21

6.1 Dynamic NGN investment and adoption models

Table 1 and Table 2 report the results of the LSDVC estimations of various NGN

investment and adoption models. The models reported in regressions (2)-(4)

represent deviations from the baseline specifications (regression (1)) as outlined in

equations (1) and (2) in terms of different selections of controls and unbundling

variables.

The coefficients of the lagged dependent variables, a

1 and c

1 , are highly

significant and substantial in all the regressions in both the investment and adoption

20

Granger causality tests require stationary time series. In order to formally test for stationarity, a “Fisher-type” (Augmented Dickey–Fuller) unit-root test, which has been designed for unbalanced panels, has been performed. This test rejects the null hypothesis that all panels contain unit roots for all the variables used in our model specifications (the results are available upon request from the authors; however, owing to the low number of observations (T = 11), the power of this test is limited). 21

Stata/IC 13.0 has been used to estimate all the regressions. Before running the regressions, a check was made on the bivariate correlations between the explanatory variables. Since two variables with high bivariate correlation produce inefficient estimates, they were excluded in the case of a higher correlation coefficient than 0.85.

~ 15 ~

models, thus indicating that the dynamic specification is correct. The coefficients

have been estimated quite precisely and are slightly larger in the adoption

regressions. This is in line with the previous literature, and suggests that consumer

inertia and switching costs are even more pronounced than adjustment costs.

As far as the unbundling price, ln(llu_price), is concerned, the coefficient estimates

are insignificant in all the regressions in Table 1 and Table 2. As indicated in section

4.2, this might be due to the low degree of variation in the unbundling price variable.

However, the variable ln(i_price_llu_sh) shows a significantly positive impact on both

NGN adoption and NGN investment. In particular, a 1% increase in the unbundling

price increases NGN adoption and NGN investment by ~0.45% and ~0.47%,

respectively.22 These results are in line with the expectations (Hypotheses 1 and 2),

as the wholesale revenue effect is explicitly controlled for by including the variable

legacy. It has also been tested whether the Eastern European countries (East = 1)

that lacked a well-developed legacy infrastructure prior to NGN deployment exhibit a

less pronounced effect of the unbundling regime. As expected (Hypothesis 4), from

the coefficient of the variable ln(i_price_llu_sh_East), it can be inferred that this effect

is offset in Eastern European countries. Accordingly, a 1% increase in the unbundling

price in regressions (4) of the coverage and adoption specifications implies an almost

null marginal increase in coverage (~ 0.1% percentage points) and adoption (0.13%).

In fact, Wald-type tests indicate that the coefficients of both linear and interactions

terms are jointly insignificant, indicating that the effect of the unbundling price is de

facto neutralized in Eastern European countries.

Moreover, the cross-price effect of the unbundling price on NGN adoption is of

particular interest, because of the lack of evidence within the existing economic

literature. Srinuan, Srinuan and Bohlin (2012) have developed an empirical

investigation to analyze direct and cross-price elasticity among different types of

broadband access technologies (xDSL, cable, fibre, mobile broadband). Data was

obtained from a random nationwide postal mail survey of Swedish households

between August and September 2009, with 2038 respondents. The results show that

the cross-price elasticity of demand for fibre, in relation to the DSL price, is 3.289. A

recent study by Grzybowski, Nitsche, Verboven and Wiethaus (2015) has used a

large database from a survey of 6446 households in Slovakia between April-July

2011 to estimate own- and cross- price elasticity of demand for different broadband

technologies (DSL, fibre, cable, WiFi and mobile broadband access). The results

show that a 1% increase in DSL price would increase the demand for fibre by

between 0.66% (at a country level) and 0.96% (at a municipality level), thus

indicating a cross-price elasticity of demand for fibre, in relation to DSL, of 0.66-0.96.

The present results on an EU level sample, which is more extensive than that of the

previous papers, are consistent with the aforementioned studies as they point out the

presence of a business migration effect from the old to the new technology

infrastructure.

22

Note that totally differentiating with respect to the unbundling price yields the constant elasticity as captured by the coefficient β2 which is independent of the unbundling market share.

~ 16 ~

As far as the competition variables are concerned, no significant pattern for

intermodal (fms) or intramodal competition (bb_ne) has been found in either type of

model. Regressions with significant estimates have indicated a negative impact of

intramodal competition on NGN coverage (regression (1)) and NGN adoption in

regression (4) and of intermodal competition with respect to NGN adoption

(regressions (2) and (4)). To the extent that these competition variables capture

market outcomes in terms of retail prices, the negative relationships can be seen as

evidence of the business migration effect. Similarly, competition stemming from the

old infrastructure (legacy) exerts a significantly negative impact on NGN investment

and adoption in almost all regression specifications. This indicates that a well-

established infrastructure also exerts a substantial replacement effect on the side of

infrastructure operators (Table 1) and substantial switching costs on the side of

consumers (Table 2). When the coefficient estimates are compared, it appears that

the replacement effect is more severe on the supply side. However, the replacement

effect is mitigated if the legacy infrastructure in Eastern European countries, as

measured by the interaction term i_legacy_East, is considered explicitly, thus again

confirming the expectations of Hypothesis 4.

As regards the cost and demand controls, the signs of all the significant coefficient

estimates are in line with the basic economic theory. Moreover, the coefficient

estimates of the main variables of interest also appear to be robust towards

alternative selections of control variables in Table 1 and Table 2. Furthermore, if the

demand and cost controls are added to the NGN coverage and NGN adoption

baseline model (“base”), the main results do not change. Overall, regressions (4) in

Table 1 and Table 2 can be considered as the final estimations as they also cover

the heterogeneity of EU member states as regards the initial conditions for NGN

deployment. Comparing these regressions, it emerges that the size of the old

broadband market, ln(bb_lines), which proxies total willingness to pay for ICT

services, has a significantly positive impact on both NGN adoption and NGN

investment. As regards the adoption model, it can also be inferred that adoption of

old broadband services, adop_bb_lines, counteracts this effect. Indeed, in the case in

which conventional broadband services enjoy broad consumer acceptance, in terms

of quality characteristics and high market saturation, the switching costs might be

substantial and hinder consumer migration to NGN services.

~ 17 ~

Table 1: Dynamic investment model (Dependent var.: ln(NGN_cov))

Regr. nr. (1) (2) (3) (4) base full Eastern_1 Eastern_2

Lag: ln(NGN_cov) 0.6415*** 0.5960*** 0.5647*** 0.5752*** (8.04) (7.58) (7.20) (7.27)

Lag: ln(llu_price) 0.1703 0.5633 0.3322 0.0854 (0.28) (0.99) (0.56) (0.15) Lag: ln(i_price_llu_sh) 0.2159** 0.2293** 0.4506*** 0.4745*** (2.05) (2.12) (2.62) (2.75) Lag: -0.3886* -0.3732* ln(i_price_llu_sh_East) (-1.71) (-1.66)

Lag: fms -0.5297 -1.1199 -0.8736 -0.6778 (-0.77) (-1.61) (-1.22) (-0.96) Lag: fms2 0.0404 0.0709 0.0570 0.0450 (0.86) (1.51) (1.19) (0.94) Lag: bb_ne 12.2182 1.9307 0.0559 -3.1084 (1.63) (0.23) (0.01) (-0.36) Lag: bb_ne2 -18.0555** -8.7270 -6.1103 -2.5497 (-2.26) (-1.01) (-0.70) (-0.29) Lag: legacy -0.0722* -0.1079** -0.1100** -0.1214*** (-1.67) (-2.35) (-2.40) (-2.65) Lag: i_legacy_East 0.1426** (2.19) Lag: urban_pop 0.0659 0.1741 0.1783 0.0903 (0.41) (1.01) (1.05) (0.51) Lag: wage -0.1920 -0.3756*** -0.4137*** -0.4735*** (-1.63) (-2.95) (-3.29) (-3.70) Lag: labcost_ict -0.0358** -0.0315** -0.0309** -0.0345** (-2.31) (-2.08) (-2.03) (-2.30) Lag: gdp -0.0000 -0.0000 -0.0000 (-1.03) (-1.45) (-1.27) Lag: ln(bb_lines) 0.6641 0.8547* 1.2482** (1.40) (1.85) (2.32) Lag: edu 0.0090*** 0.0106*** 0.0100*** (2.62) (3.10) (2.95) Lag: nri -1.2984 -1.3285 -1.1536 (-1.47) (-1.50) (-1.34) Year dummies YES YES YES YES

R2 (within) 0.8213 0.8364 0.8431 0.8478 AR(2) (p-value) 0.301 0.274 0.260 0.384 Observations 178 178 178 178

The LSDVC standard errors in regressions (1)-(4) have been bootstrapped with bias correction initialized by the Arellano and Bond estimator (Arellano & Bond, 1991) for estimates. Note that there are no standard post-estimation tests available in STATA for the user written “xtlsdvc” command (Bruno, 2005b). Therefore, the R

2 within has been provided on the basis of an LSDV regression with a

lagged dependent variable. Moreover, a specification test, based on the Arellano-Bond test for autocorrelation in the residuals, has also been provided. If the assumption of serial independence in the original errors, ε´s and φ´s, is correct, the transformed residuals should not show any significant AR(2) test statistics. The t-statistics are reported in parentheses; * p < 0.10, ** p < 0.05, *** p < 0.01

~ 18 ~

Table 2: Dynamic adoption model (Dependent var.: ln(NGN_adop))

Regr. nr. (1) (2) (3) (4) base fms full Eastern

Lag: ln(NGN_adop) 0.6533*** 0.6464*** 0.6431*** 0.6270*** (12.32) (11.38) (11.35) (12.34) Lag: ln(llu_price) -0.4199 -0.3805 -0.6332 -0.6661 (-0.97) (-0.87) (-1.46) (-1.42) Lag: ln( i_price_llu_sh) 0.2877*** 0.2987*** 0.2920*** 0.4488*** (3.85) (4.17) (4.15) (4.26) Lag: -0.3147** ln(i_price_llu_sh_East) (-2.12)

Lag: fms -0.4413 -0.2642** -0.1513 -0.1824* (-1.10) (-2.41) (-1.31) (-1.66) Lag: fms2 0.0122 (0.43) Lag: bb_ne 1.2408 -0.6535 4.7136 -3.7940*** (0.21) (-0.11) (0.79) (-2.78) Lag: bb_ne2 -5.4528 -3.3921 -9.0910 (-0.96) (-0.61) (-1.59) Lag: legacy -0.0578** -0.0421* -0.0437* -0.0337 (-2.30) (-1.87) (-1.80) (-1.35) Lag: gdp -0.0000 -0.0000 -0.0000 -0.0000 (-0.95) (-1.32) (-1.56) (-1.46) Lag: ln(bb_lines) 0.5685* 0.3020 0.4584 1.1270** (1.73) (0.68) (1.06) (2.46) Lag: adop_bb_lines -3.0634* (-1.76) Lag: edu 0.0863*** 0.0729** 0.0347 0.0011 (2.91) (2.26) (0.98) (0.43) Lag: nri 0.0109 -0.6091 -0.1894 -0.2750 (0.04) (-1.20) (-0.36) (-0.53) Lag: labcost_ict -0.0233** -0.0263*** (-2.44) (-2.95) Lag: urban_pop 0.1770 0.2148* (1.59) (1.95) Year dummies YES YES YES YES

R2 (within) 0.8625 0.8695 0.8770 0.8810 AR(2) (p-value) 0.376 0.714 0.692 0.631 Observations 196 196 196 196

The LSDVC standard errors in regressions (1)-(4) have been bootstrapped with bias correction initialized by the Arellano and Bond estimator (Arellano & Bond, 1991) for estimates. Note that there are no standard post-estimation tests available in STATA for the user written “xtlsdvc” command (Bruno, 2005b). Therefore, the R

2 within has been provided on the basis of an LSDV regression with a

lagged dependent variable. Moreover, a specification test, based on the Arellano-Bond test for autocorrelation in the residuals, has also been provided. If the assumption of serial independence in the original errors, ε´s and φ´s, is correct, the transformed residuals should not show any significant AR(2) test statistics. The t-statistics are reported in parentheses; * p < 0.10, ** p < 0.05, *** p < 0.01

~ 19 ~

6.2 The NGN take-up rate model

As described in Section 3, the take-up rate relates NGN adoption to NGN coverage.

This index is extremely important for three reasons: first, it is a relevant indicator of

the willingness of consumers to migrate to a new infrastructure; second, it is a

measure of capacity utilization; third, and perhaps more important, it is a key policy

variable defined by the EC in its DAE targets. The results reported in Section 6.1 on

coverage and adoption show that both NGN coverage and adoption is positively

affected by an increase in the access price to the old networks, thus implying that the

expected effect of this price on the take-up rate is ex ante unclear. The main results

of the static NGN take-up rate model are reported in Table 3. The F-test (F_f), at the

bottom of Table 3, shows that country-level fixed-effects are highly significant, which

in turn implies that pooled OLS would produce inconsistent estimates if the fixed-

effects were correlated to the independent variables.23 Wooldridge´s test for serial

correlation in panel data (Wooldridge, 2002) clearly indicates the presence of a first-

order serial correlation (e.g. F(1, 24) = 34.074 for the baseline model in regression

(1)). This test is robust to conditional heteroscedasticity, which is present in the take-

up rate model specifications. Accordingly, two-way fixed-effects regressions have

been employed with an AR(1) disturbance in regressions (1)-(4). It should be noted

that all the demand and cost side controls have been included in regressions (1)-(4),

as outlined in Section 5.2.

The results show that an increase in the local loop unbundling price, llu_price, has a

significantly negative impact on the take-up rate of NGN connections. Putting

together the results from the previous Sections, the overall results show that both

adoption and coverage do in fact increase with an increase in the regulated access

price, but the effects on coverage slightly dominates the effect on the demand side

and hence reduces the take-up rate (Hypothesis 3). This effect, though weakened

by the simultaneous effects on adoption and coverage, is constant across the

specifications. Accordingly, an increase in the unbundling price by 1€, increases the

NGN take-up rate by ~ 1 percentage point. Similarly, and in line with our previous

results, controlling for the presence of Eastern countries (regression 4) does not

affect the results, which means that the role of the unbundling price is irrelevant in

those countries to sustain NGN take-up.

Most cost and demand side controls do not seem to play any relevant role, while

fixed-mobile substitution does. The more intense the intermodal competition is, the

lower the NGN take-up rate; this effect is also non-linear, as suggested by the model

specifications in regressions (3)-(4). The corresponding coefficients (regression (4))

on the fms and fms2 variables point to an inverted U-shaped relationship, with an

optimal level of competition intensity for fms ~ 6.18, which is well above the grand

mean value (fms ~ 3.37 (Table A.2)). Hence, on average intermodal competition from

mobile networks exerted a positive impact on the NGN take-up rate in the past.

23

A robust Hausman test clearly rejects the random effect model assumption (the Sargan-Hansen test results are significant at the 1% level; not reported here, but available upon request). Clearly, the EU27 member states do not represent a random sample drawn from the population of all countries.

~ 20 ~

6.3 Further robustness tests

This section presents additional estimations that can be used to examine the

robustness of the main results. The robustness tests refer to (i) an alternative

estimator (fixed-effects instrumental variable (IV) estimation in regressions (1)-(3))

and (ii) an alternative specification of the dependent variable (ln(NGN_gap)24 in

regression (4)).

As described in section 4.2, the unbundling price has been instrumented with the

price of shared access, sa_price, as well as with some other (excluded) exogenous

cost shifters (population density, pop_dens, and the long-term interest rate, lt_ir). In

line with the previous sections, we employ the bias-corrected LSDVC estimator for

the dynamic NGN adoption and coverage specification in regressions (1)-(2) and the

ordinary (LSDV) fixed-effect estimator for the static NGN take-up rate and NGN gap

model in regressions (3)-(4).

A first stage regression shows that the instruments are jointly highly significant (F =

48.07). From regression (1)-(2) in Table 4, the main estimation results carry over

quite well as regards the dynamic NGN coverage (regression (1)) and NGN adoption

(regression (2)) models, where the “full” model specifications have been re-

estimated, as reported in Table 1 and Table 2. The same results on the unbundling

price also show up in the IV NGN take-up rate (regression (3)) model. Focusing on

the role of the unbundling price we re-estimated the structure of the “ull” model, as

reported in regression (2) of Table 3. Whereas the main term is now insignificant, the

interaction term picks-up the negative relation which is significant at the 1% level.

As far as the ln(NGN_gap) model is concerned, supportive evidence has also been

found on the impact of the unbundling price. The positive coefficient estimate can

now be expected, in view of the construction of the variable ln(NGN_gap). This also

holds for the other explanatory variables, and, in particular, the market size now

exerts a positive and significant effect.

24

In order to normalize the series, logs of the dependent variable were considered.

~ 21 ~

Table 3: Static take-up rate model (Dependent var.: NGN_tur)

Regr. nr. (1) (2) (3) (4) base llu fms Eastern

Lag: llu_price -0.0097* -0.0096* -0.0102* -0.0116* (-1.77) (-1.74) (-1.88) (-1.81) Lag: i_price_llu_sh -0.0013 (-0.11) Lag: i_llu_price_East 0.0036 (0.35) Lag: fms -0.0280* -0.0283* 0.1119* 0.1186* (-1.68) (-1.69) (1.68) (1.76) Lag: fms2 -0.0094** -0.0096** (-2.14) (-2.17) Lag: bb_ne -0.9742 -0.9714 -0.9342 -1.0172 (-1.22) (-1.20) (-1.19) (-1.31) Lag: bb_ne2 1.2196 1.2210 1.1889 1.2858 (1.49) (1.48) (1.48) (1.61) Lag: legacy 0.0021 0.0021 0.0051 0.0064 (0.49) (0.49) (1.12) (1.35) Lag: gdp 0.0000 0.0000 -0.0000 0.0000 (0.21) (0.22) (-0.04) (0.57) Lag: ln(bb_lines) 0.1502 0.1501 0.1667 0.1779 (1.13) (1.12) (1.26) (1.31) Lag: adop_bb_lines -0.3135 -0.3169 -0.2531 -0.2495 (-1.00) (-1.00) (-0.81) (-0.80) Lag: edu -0.0000 0.0000 -0.0003 -0.0005 (-0.00) (0.00) (-0.05) (-1.15) Lag: nri 0.0799 0.0794 0.1069 0.1072 (1.16) (1.14) (1.53) (1.55) Lag: labcost_ict -0.0001 -0.0001 -0.0004 -0.0004 (-0.10) (-0.09) (-0.24) (-0.29) Lag: urban_pop 0.0364 0.0366 0.0409 0.0282 (1.47) (1.47) (1.62) (1.04) Lag: wage 0.0271* 0.0275* 0.0287* 0.0346** (1.83) (1.82) (1.92) (2.18) Constant -1.0575*** -1.0650*** -1.0148*** -1.0273*** (-2.91) (-2.89) (-2.89) (-2.86) Year dummies YES YES YES YES

R2 (within) 0.2368 0.2378 0.2518 0.2547 F 2.0504 1.9629 2.1170 2.0501 F_f 2.6233 2.6220 2.4251 2.6249 Observations 200 200 200 200 Note that panel-by-panel Cochrane-Orcutt method decreases the number of maximum observations by the number of available groups. In addition, some values in our panel data set are missing, as pointed out in section 4. All regressions include country fixed effects and period effects. The t-statistics are reported in parentheses; * p < 0.10, ** p < 0.05, *** p < 0.01

~ 22 ~

Table 4: Robustness regressions (Dependent var.: (regr. (1): ln(NGN_cov); regr. (2): ln(NGN_adop); regr. (3): NGN_tur; regr. (4): ln(NGN_gap)

Regr. nr. (1) (2) (3) (4) cov_full_IV

_LSDVC adop_full_IV

_LSDVC tur_ull_IV

_FE_robust ln(NGN_gap) _FE_AR(1)

Lag: Dependent var. 0.6362*** 0.6534*** (8.50) (11.37) Lag: ln(llu_price) 0.1620 -0.4045 (0.33) (-0.92) Lag: llu_price 0.0083 0.1195*** (1.33) (2.97) Lag:ln(i_price_llu_sh) 0.1538* 0.2995*** -0.0192*** (reg (3): i_price_llu_sh) (1.66) (4.08) (-3.22)

Lag: fms -1.0029* -0.2125* -0.0426* -0.0562 (-1.82) (-1.82) (-1.74) (-0.29) Lag: fms2 0.0595 (1.58) Lag: bb_ne 3.8542 2.4236 -2.2000* -1.5749 (0.48) (0.38) (-1.96) (-0.34) Lag: bb_ne2 -8.8926 -6.3057 2.7773** 2.5772 (-1.04) (-1.03) (2.56) (0.50) Lag: legacy -0.1090*** -0.0454* 0.0074 -0.0229 (-3.02) (-1.81) (1.35) (-0.58) Lag: gdp -0.0000 -0.0000 0.0000 -0.0000 (-0.40) (-1.43) (0.89) (-0.68) Lag: ln(bb_lines) 0.5454 0.4030 -0.0029 1.4941** (1.12) (0.86) (-0.05) (2.15) Lag: adop_bb_lines -0.5342* -1.8500 (-1.94) (-0.80) Lag: edu 0.0064* 0.0580 0.0086 -0.0498 (1.88) (1.61) (1.01) (-0.81) Lag: nri -1.0308 -0.4300 -0.0365 0.2883 (-1.39) (-0.78) (-0.72) (0.47) Lag: urban_pop 0.1829 0.1562 0.0153 0.4764 (1.28) (1.27) (0.63) (1.24) Lag: wage -0.2973*** 0.0297** -0.3824** (-2.76) (2.41) (-2.26) Lag: labcost_ict -0.0149 -0.0096 0.0012 -0.0200 (-1.21) (-0.82) (0.54) (-1.41) Constant -1.6167 -29.5569*** (-0.79) (-3.23) Year dummies YES YES YES YES

R2 (within) 0.7795 0.7410 0.4032 0.1555 F_f 12.27 8.23 7.8381 1.12 AR(2) (p-value) 0.804 0.600 Observations 200 196 218 216 Estimates in regressions (1)-(3) are based on the two-stage least squares IV estimator. Regression (4) is based on panel-by-panel Cochrane-Orcutt method to eliminate first-order serial correlation. All regressions include country fixed effects and period effects. The t-statistics are reported in parentheses and are bootstrapped in regressions (1)-(2) and robust to arbitrary forms of heteroscedasticity and serial correlation in regression (3); * p < 0.10, ** p < 0.05, *** p < 0.01

~ 23 ~

7 Summary and Conclusions

The aim of this paper has been to provide evidence on a hotly debated issue, i.e.

how to simultaneously incentivize the adoption and the investment in next generation

broadband technology. In particular, the focus of the paper has been on the potential

role of the access price on the old broadband infrastructure, which is set directly by

NRAs and acts as a key policy variable to speed up investment and the adoption of

new ultra-fast broadband connections.

Results show that NGN coverage and adoption are characterized by the presence of

path dependency: this implies that policies aimed at fostering retail migration are

important to sustain demand expansion. At the same time, the existing access price

regulation, i.e. the LLU price, could affect NGN adoption indirectly, albeit

considerably. The data show that relaxing the LLU regulation, i.e. allowing an

increase in access prices for the old legacy infrastructures, could help to support a

demand expansion and reduce the price differentials between the prices of standard

broadband services and the NGN-based ones. However, we found that there is

considerable heterogeneity among EU member states implying, in particular, that the

impact of unbundling policies are strongly weakened in Eastern European countries,

where the regulated old broadband infrastructures are much less developed.

Furthermore, the effect of an increase in the LLU access price is greater for NGN

coverage than for adoption, thus widening the gap between adoption and coverage

and therefore reducing the take-up rate. In other words, although it positively affects

NGN adoption and NGN coverage, an increase in LLU prices could also generate

extra-capacity without enhancing sufficient ultra-fast broadband demand, thus

implying that, on the demand side, additional policies are needed to sustain demand

expansion. This result is reminiscent of Tinbergen´s maxim according to which the

number of policy instruments must be equal to the number of policy targets.

Consequently, in order to achieve the mid-term dual DAE goals, both the demand

and the supply sides of the European broadband markets need to be stimulated.

Significant investments in telecom and/or cable infrastructure are needed on the

supply side in order to enable much higher internet speeds. Instead, on the demand

side, the consumers need to be persuaded about the potential benefits of new

applications that make use of these higher speeds and need to be offered affordable

prices in order to subscribe, e.g. via vouchers, tax deductions or other public demand

stimuli. Only on the assumption that development of content and applications will

autonomously evolve sufficient demand after the necessary infrastructure has

already been put in place and the welfare loss due to slower migration is not too

large, the negative impact of the access price on the take-up rate can be considered

as a second-order effect.

~ 24 ~

Acknowledgements

We would like to thank Robert Albon, Jan Bouckaert and Michal Grajek for their

suggestions on the previous versions of this paper, as well as the participants in the

seminar at the University of Antwerp and DG Connect (European Commission,

Brussels).

~ 25 ~

Annex

Tables A.1, A.2, A.3 and Figures A.1 and A.2.

Table A.1: Description of the variables and sources

Variable Description Source

Dependent variables

NGN coverage NGN_cov (household weighted)

Total number of homes passed by FTTx technologies (Fibre-to-the-home; Fibre-to-the-building; Fibre-to-the-curb; Fibre-to-the-last amplifier/DOCSIS 3.0). “Homes passed” refers to the total number of premises. “Premises” is a home or place of business, normalized to each country’s total number of households.

FTTH Council

Europe

Euromonitor

(households)

NGN adoption NGN_adop (household weighted)

Total number of subscribers in terms of “homes connected” by FTTx technologies. “Subscribers” refers to premises that uses at least one service in this connection under a commercial contract, normalized to each country’s total number of households.

FTTH Council

Europe

NGN take-up rate,

NGN_tur

Ratio between NGN adoption and NGN coverage. FTTH Council

Europe

NGN gap NGN_gap

Difference between NGN coverage and NGN adoption.

FTTH Council

Europe

Main explanatory variables: Regulation

Average total cost of the full LLU, llu_price

Monthly average total cost of the full LLU in €. EU Digital Agenda

Scoreboard

Average cost of shared access, sa_price

Monthly average total cost of shared access in €. EU Digital Agenda

Scoreboard

Main explanatory variables: Competition

Entrant's market share, bb_ne

New entrant's retail market share in fixed broadband lines.

Communication

Committee

(COCOM)

Mobile-to-fixed ratio, fms

Ratio of Mobile Lines to Fixed Lines (Absolute). Market Line Extract

Fixed legacy, legacy

Total number of active fixed landlines per 100 inhabitants. An active line connects the subscriber’s terminal equipment to the public switched telephone network PSTN lines.

ITU

Share of LLU lines, ms_llu

Share of unbundled local loop lines to the total retail broadband lines.

EU Digital Agenda

Scoreboard

~ 26 ~

Table A.1 ctd.

Demand control variables

Broadband lines, bb_lines

Number of total retail broadband connections based on DSL and coax cable that enable a higher than 144 Kbit/s download speed but exclude FTTx lines.

EU Digital Agenda

Scoreboard

Broadband adoption, adop_bb_lines

Number of total broadband connections adopted by consumers divided by total population.

EU Digital Agenda

Scoreboard

Networked Readiness Index, nri

Propensity of a country to exploit the opportunities offered by information and communication technology (ICT).

Euromonitor

Education, edu

Percentage of population having attained secondary or higher education, for the population aged 25 to 64 years.

Eurostat

GDP per capita, gdp

GDP per capita (total) and PPP adjusted to current US$.

World Bank

Euromonitor

(population)

Cost control variables

Hourly wage, wage

The manufacturing wage per hour in € and current prices with fixed 2012 exchange rates.

Euromonitor

Labour cost, labcost_ict

Annual labour cost index for the Information and Communication branch by NACE Rev. 2 normalized to 100 in 2008. The index measures the development of the total cost, on an hourly basis, to employ the labour force, and it includes wages and salaries, social security contributions and taxes, but excludes subsidies.

Eurostat

Urban population, urban_pop

Population of a country that lives in an urban environment as a percentage of the total population.

MarketLine

Population density pop_dens

Population density in number of inhabitants per Square Kilometre.

Market Line Extract

Long-term interest rate, lt_ir

Long-term interest rate for debt security issued after 10 years of maturity at the local currency unit rate.

European Central Bank

~ 27 ~

Table A.2: Summary statistics

Variable Obs Mean Std. Dev. Min Max

NGN_adop 296 0.0730204 0.1090571 0 0.5471706

NGN_cov 296 0.3830631 0.4678066 .4678066 2.306572

NGN_tur 226 0.2199841 0.1641123 0.0000517 0.7222222

NGN_gap 296 0.2651446 0.2945837 0 0.9983765

llu_price 266 11.45305 4.303125 5.28 42

sa_price 266 5.397406 3.431645 0.74 23.89

fms 270 3.371881 1.667801 1.2819 10.9396

bb_ne 267 0.501393 0.1558175 0 1

legacy 270 40.41304 13.08719 13.86 66.38055

bb_lines 267 3723236 5769546 13738 27960396

ms_llu 266 0.1064223 0.1461762 0 0.6772212

nri 270 4.578519 .6294371 3.2 6

gdp 270 30200.01 13641.82 8730.803 90789.65

edu 270 73.53926 16.01936 23.6 93.4

wage 270 11.05556 7.861194 0.8 38.7

urban_pop 270 72.43043 11.89043 49.4118 97.4945

labcost_ict 270 99.84741 15.33449 47.9 163.5

pop_dens 270 174.247 237.3405 17.1923 1285.241

lt_ir 296 4.50125 2.227483 0.22 22.5

~ 28 ~

Table A.3: Direct Granger-causality tests

Since Granger-causality tests include several lags of the right-hand side variable,

including the lagged dependent variable, the Arellano and Bond (1991) difference-in-

difference GMM estimator has been employed with a maximum number of three lags

of the right-hand side variables and internal instruments. Period effects have been

included. Granger-causality tests are Wald tests of the joint significance of the

respective coefficients which are χ2 distributed. The standard errors have been

adjusted for clustering within countries and are robust to heteroscedasticity; the p-

values are reported.

GMM (NGN adoption) p-value H0: β1, β2 = 0 Answer

Does the LLU price cause NGN adoption?

Does NGN adoption cause LLU price?

0.0077 0.2890

Rejected Not Rejected

Yes No

Do new BB entrants market share cause NGN adoption? Does NGN adoption cause new BB entrants market share?

0.2078 0.7779

Not Rejected Not Rejected

No No

Does fixed to mobile substitution rate cause NGN adoption? Does NGN adoption cause fixed to mobile substitution rate?

0.1365 0.7434

Not Rejected Not Rejected

No No

Does the no. of active fixed landlines cause NGN adoption? Does NGN adoption cause the no. of active fixed landlines?

0.0632 0.7600

Rejected Not Rejected

Yes No

GMM (NGN coverage) p-value H0: β1, β2 = 0 Answer

Does the LLU price cause NGN coverage?

Does NGN coverage cause LLU price?

0.0316 0.7613

Rejected Not Rejected

Yes No

Do new BB entrants market share cause NGN coverage? Does NGN coverage cause new BB entrants market share?

0.3365 0.8522

Not Rejected Not Rejected

No No

Does fixed to mobile substitution rate cause NGN coverage? Does NGN coverage cause fixed mobile substitution rate?

0.0011 0.6111

Rejected Not Rejected

Yes No

Does the no. of active fixed landlines cause NGN coverage? Does NGN coverage cause the no. of active fixed landlines?

0.0066 0.5930

Rejected Not Rejected

Yes No

GMM (NGN take-up rate) p-value H0: β1, β2 = 0 Answer

Does the LLU price cause NGN take-up rate?

Does the NGN take-up rate cause LLU price?

0.0865 0.7138

Rejected Not Rejected

Yes No

Do new BB entrants market share cause NGN take-up rate? Does the NGN take-up rate cause new BB entrants market share?

0.9207 0.9941

Not Rejected Not Rejected

No No

Does the FMS rate cause NGN take-up rate? Does the NGN take-up rate cause FMS rate?

0.9986 0.9967

Not Rejected Not Rejected

No No

Does the no. of active fixed landlines cause NGN take-up rate? Does the NGN take-up rate cause the no. of active fixed landlines?

0.8553 0.6951

Not Rejected Not Rejected

No No

~ 29 ~

Figure A.1: NGN coverage, adoption and take-up rates in the average EU member state (Source: FTTH Council Europe)

~ 30 ~

Figure A.2: NGN adoption, coverage and take-up: EU15 vs. Eastern European countries (Source: FTTH Council Europe)

~ 31 ~

References

Arellano, M. and S. Bond (1991), Some Tests of Specification for Panel Data: Monte

Carlo Evidence and an Application to Employment Equations, Review of

Economic Studies 58, 277-297.

Bacache, M., Bourreau, M. and G. Gaudin (2014), Dynamic Entry and Investment in

New Infrastructures: Empirical evidence from the Fixed Broadband Industry,

Review of Industrial Organization 44, 179-209.

Bouckaert, J., van Dijk, T. and F. Verboven (2010), Access Regulation, Competition,

and Broadband Adoption: An International Study, Telecommunications Policy

34, 661-671.

Bourreau, M., Cambini, C. and P. Dogan (2012), Access Pricing, Competition, and

Incentives to Migrate From “Old” to “New” Technology, International Journal of

Industrial Organization 30, 713-723.

Bourreau, M., Cambini, C. and P. Dogan (2014), Access Regulation and the

Transition from Copper to Fiber Networks in Telecoms, Journal of Regulatory

Economics 45(3), 233- 258.

Bresnahan, T. and M. Trajtenberg (1995), General Purpose Technologies: ‘Engines

of Growth’?, Journal of Econometrics 65, 83-108.

Briglauer, W. (2015), How EU sector-specific regulations and competition affect

migration from old to new communications infrastructure: recent evidence from

EU27 member states, Journal of Regulatory Economics 48(2), 194-217.

Briglauer, W. (2014), The Impact of Regulation and Competition on the Adoption of

Fiber-Based Broadband Services: Recent Evidence from the European Member

States, Journal of Regulatory Economics 46, 51-79.

Bruno, G. (2005a), Approximating the bias of the LSDV estimator for dynamic

unbalanced panel data models, Economics Letters 87, 361-366.

Bruno, G. (2005b), Estimation and inference in dynamic unbalanced panel-data

models with a small number of individuals, The Stata Journal 5, 473-500.

Cambini, C. and Y. Jiang (2009), Broadband Investment and Regulation: A Literature

Review, Telecommunications Policy 33, 559-574.

Czernich, N., Falck, O., Kretschmer, T. and L. Wößmann (2011), Broadband

Infrastructure and Economic Growth, The Economic Journal 121, 505-532.

~ 32 ~

Distaso, W., Lupi, P. and F.M. Maneti (2006), Platform competition and broadband

uptake: Theory and empirical evidence from the European union, Information

Economics and Policy 18, 87-106.

ERG – European Regulators Group (2007), Opinion on Regulatory Principles of

NGA, 16rev2, Brussels, available at:

http://erg.ec.europa.eu/doc/publications/erg07_16rev2_opinion_on_nga.pdf.

European Commission (2000), Regulation (EC) No 2887/2000 of the European

Parliament and of the Council of 18 December 2000 on unbundled access to

the local loop, Official Journal of the European Communities, Brussels,

available at: http://eur-

lex.europa.eu/LexUriServ/LexUriServ.do?uri=OJ:L:2000:336:0004:0008:EN:PD

F.

European Commission (2002a), Directive 2002/21/EC of the European Parliament

and of the Council of 7 March 2002 on a common regulatory framework for

electronic communications networks and services, Framework Directive,

Brussels.

European Commission (2002b), Directive 2002/19/EC of the European Parliament

and of the Council of 7 March 2002 on access to, and interconnection of,

electronic communications networks and associated facilities, Access Directive,

Brussels.

European Commission (2010), A Digital Agenda for Europe, COM(2010) 245,

Brussels.

European Commission (2014), Broadband Coverage in Europe 2013: Mapping

progress towards the coverage objectives of the Digital Agenda. Study prepared

for the European Commission DG Communications Networks, Content &

Technology by IHS and VVA consulting, Brussels.

FTTH Council Europe (2013), FTTH Business Gudie, Edition 4, Business Committee,

available at:

http://www.ftthcouncil.eu/documents/Publications/FTTH_Business_Guide-

_2013_V4.0.pdf.

FTTH Council Europe (2014), FTTH Handbook, Edition 6, D&O Committee, available

at: http://www.ftthcouncil.eu/EN/home/forms/form-handbook.

~ 33 ~

Grajek, M., and L.H. Röller (2011), Regulation and Investment in Network Industries:

Evidence from European Telecoms, Journal of Law and Economics 55, 189-

216.

Granger, C.W.J. (1969), Investigating causal relations by econometric models and

cross-spectral methods, Econometrica 37, 424-438.

Grzybowski L., Nitsche R., Verboven. F. and L. Wiethaus (2015), Market Definition

for Broadband Internet in Slovakia. Are Fixed and Mobile Technologies in the

Same Market?, Information Economics and Policy.

Inderst, R. and M. Peitz (2012), Market Asymmetries and Investments in NGA,

Review of Network Economics 11(1), Art. 2.

Kiiski, S. and M. Pohjola (2002), Cross-country diffusion of the Internet, Information

Economics and Policy 14, 297-310.

Lee, S. and S. Lee (2010), An empirical study of broadband diffusion and bandwidth

capacity in OECD countries, Communications & Convergence Review 2(1).

Lee, S., Marcu, M. and S. Lee (2011), An empirical analysis of fixed and mobile

broadband adoption, Information Economics and Policy 23, 227-233.

Minamihashi, N. (2012), Natural Monopoly and Distorted Competition: Evidence from

Unbundling Fiber-Optic Networks, Working paper no.2012-26, Bank of Canada,

available at: http://www.bankofcanada.ca/wp-content/uploads/2012/08/wp2012-

26.pdf.

Nardotto, M., Valletti, T. and F. Verboven (2015), Unbundling the incumbent:

Evidence from UK broadband, Journal of the European Economic Association

forthcoming.

Nickell, S. (1981), Biases in dynamic models with fixed effects, Econometrica 49,

1417-1426.

Röller, L.H. and L. Waverman (2001), Telecommunications Infrastructure and

Economic Development: A Simultaneous Approach, American Economic

Review 91, 909-923.

Samanta, S.K., Martin, R., Guild, K. and H. Pan (2012), The Diffusion of High Speed

Broadband: A Cross Country Analysis, working paper, available at:.

http://ssrn.com/abstract=1997113.

~ 34 ~

Schmutzler, A. and D. Sacco (2011), Is there a U-shaped relation between

competition and investment?, International Journal of Industrial Organization 29,

65-73.

Srinuan P., Srinuan C. and E. Bohlin (2012), Fixed and Mobile Broadband

Substitution in Sweden, Telecommunications Policy 36, 237-251.

Wallsten, S. and S. Hausladen. (2009), Net Neutrality, Unbundling, and Their Effects

on International Investment in Next-Generation Networks, Review of Network

Economics 8, Art. 6.

Wooldridge, J. (2002), Econometric Analysis of Cross Section and Panel Data, MIT

Press, Cambridge.